AI Incorporates Appetites of Giants in Network Switching: Ethernet or Infiniband?

The switch market for data processing centers (data center) will grow by 50% due to the growing demand for artificial intelligence. According to forecasts by the analyst Dell’oro, Mamha Budzhelben, the demand for AI will increase the volume of this market double by 2027.

Currently, the share of the market of the network switches in the total volume of the market is less than 10%. At the same time, about 90% of AI deployment use Infiniband technology, not Ethernet. Thanks to this, Nvidia took second place in the switching market after Cisco.

In the AI segment, key factors are throughput and delay. Infiniband is characterized by low delays and losses of packages compared to Ethernet. And this is critically important for expensive and long processes of teaching AI.

However, Infiniband is inferior to Ethernet in the maximum passing strip. The latest QUANTUM Infiniband switches from NVIDIA have a capacity of 25.6 Tbit/s and ports of 400 Gbit/s. And Ethernet has already reached 51.2 tbit/s and 800 Gb/s, respectively.

At the same time, the usual server does not need such high-speed ports. But AI servers are equipped with a 400 gigabit network adapter for each GPU, and the GPU can be 4-8 pieces. Such AIs are filled with a huge bandwidth.

To compete with Infiniband, Ethernet even appeared technologies for managing loss and delays in packages. This will allow Ethernet to take 20% of the switching market for AI by 2027 according to Dell’oro forecasts. The main growth drivers will be hyper-scale data processing centers and cloud providers.

By 2025, most of the switches deployed for AI will have ports of 800 Gbit/s. And by 2027, their performance can be